How Much of "Cracking the Code of the Rohonc Codex" Can Be Checked from the Open Facsimile? A Claim-to-Folio Audit with a Pre-Coding Frozen Plan

Research manuscript, version 0.9.1 (3 October 2026). It is version 0.8 (deposited with Zenodo version 1.2.0) with journal-format changes (page numbers, labelled abstract parts, a Conclusions section that repeats statements made in the text, and a Competing interests heading) and two reporting corrections: the earlier statement that the data deposit contains no transcription is narrowed to describe the Unicode sign strings in its evidence notes, and the facsimile-resolution figure of about 110 dpi per page, which could not be reproduced from its recorded derivation, is removed. It has not been peer reviewed. The data, claim list, codebook, folio concordance and code are deposited separately: https://doi.org/10.5281/zenodo.22883816 (version 1.0.0, original data); all versions: https://doi.org/10.5281/zenodo.22883815. Király and Tokai (2018) proposed a systematic decipherment of the Rohonc Codex (Library of the Hungarian Academy of Sciences, MS K 114) and anchored their argument in dozens of concrete observations about specific folios: pencil foliations, incomplete lines, text running inside picture frames, verbatim parallels between passages, and numeral sequences. This study asks a narrow, non-interpretive question: what proportion of those folio-anchored observations can a reader using only the article and the openly downloadable facsimile reproduce from that facsimile, the low-resolution monochrome PDF served by the Academy's repository, without making any sign-identification decision that the article does not itself publish? Thirty checkable claims from the article's sections I–III were listed as excerpts or abridged transcriptions with page locators and frozen, together with a four-class codebook (exactly reproducible / partly / dependent on an unpublished sign decision / not found), before any claim was coded. In the stored AI-coded results for the frozen thirty-claim list, fifteen of the thirty claims (0.50; Wilson 95 % interval 0.33–0.67) were classified as exactly reproducible, fourteen as partly reproducible, one as not checkable because the article gives no locator for one side of a claimed parallel, and none as dependent on an unpublished sign decision. These classifications have not been independently validated by a human. This finding is limited to the thirty retained claims and does not extend to the excluded candidates or to the article as a whole. In the AI-coded records, facsimile resolution is recorded as limiting thirteen of the thirty verdicts and as the sole reason for ten of the fourteen partial verdicts; these are recorded attributions, not independently established causes. The descriptive audit does not assess the proposed decipherment or establish causal effects. It reports AI-coded traceability judgments and locators at which higher-resolution access could permit reassessment. The claim list, codebook, folio concordance and code are deposited openly; no manuscript images are redistributed. AI use: Claude, through Claude Code, selected the claims, prepared the codebook, carried out both coding passes, and prepared the manuscript and release materials, including versions 0.8 and 0.9. OpenAI Codex was used for the 23 September 2026 consistency checks and editorial changes and performed the historical Zenodo and OSF operations; Perplexity and ChatGPT performed external audits. Version 0.9.1 corrects two statements identified by an external audit with ChatGPT; the corrections were specified by Claude through Claude Code and applied with OpenAI Codex. Veronika Lux initiated the project, made the source-access and publication-venue decisions, initiated external research runs and gave approvals. She did not code the claims and did not check the coding claim by claim. These recorded roles do not establish independent human validation. The full AI Use Disclosure Statement, with recorded model identifiers and the author's responsibility statement, is in the manuscript.